Automated in situ optimization and disorder mitigation in a quantum device

Technical Summary: Automated In Situ Optimization and Disorder Mitigation in a Quantum Device

Problem Statement
The tuning of quantum devices, particularly those requiring the adjustment of multiple parameters to reach a desired operating configuration, is often hindered by material imperfections and unpredictable disorder. These imperfections can obscure the navigation through large parameter spaces, making it difficult to manually achieve optimal device performance. Specifically, in Quantum Point Contacts (QPCs), disorder and scattering tend to smooth out and distort the characteristic quantized conductance steps (plateaus separated by sharp jumps), which are fundamental hallmarks of quantum physics and essential for applications like spin qubit readout and interferometry. The challenge lies in finding a tunable potential landscape that can restore these "staircase" signatures despite the presence of microscopic disorder.

Methodology
The authors implement an automated evolutionary strategy to optimize the potential landscape of a QPC defined within a GaAs-based two-dimensional electron gas (2DEG). The device features a $3 \times 3$ array of electrostatic top-gates (pixel gates) patterned atop the constriction, alongside outer split-gates.

• Algorithm: The study utilizes the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). This gradient-free, stochastic algorithm is chosen for its simplicity, robustness against noise, and insensitivity to the specific details of the loss landscape.
• Optimization Metric (Loss Function): A custom loss function $L$ is defined to quantify how "step-like" the conductance curve is as the channel is constricted. The function sums the cube roots of the differences between adjacent conductance values, favoring a few large jumps (steps) over many small variations. It includes a penalty for completely closing the QPC and a small offset to mitigate noise, but crucially, it contains no prior knowledge of the specific integer quantized values or the underlying disorder distribution.
• Simulation Approach: Initial optimization was performed using the KWANT software package to solve tight-binding models. To improve robustness, the optimization was not performed on individual pixel voltages directly but on their Fourier modes ($A_{\alpha\beta}$). The average mode ($A_{00}$) served as the sweeping parameter ($V_{QPC}$) to constrict the channel, while other modes were optimized. Disorder was modeled as a random onsite potential with a defined correlation length.
• Experimental Approach: The same CMA-ES routine was applied to a physical device fabricated on a GaAs heterostructure. Due to fabrication constraints (specifically, the need for electrical routing that broke device symmetry), the experimental protocol was adapted: the sweeping parameter $V_{QPC}$ was applied to two opposing pixels, while the algorithm freely optimized the remaining seven pixel voltages. Measurements were conducted in a dilution refrigerator at millikelvin temperatures with an out-of-plane magnetic field.

Key Contributions

1. Demonstration of Automated Tuning: The paper presents a proof-of-principle for fully automated, in situ optimization of a quantum device using an evolutionary strategy without human intervention.
2. Disorder Mitigation: The study demonstrates that the algorithm can adapt to random disorder potentials. By optimizing the gate voltages, the system can recover clear conductance plateaus that were previously degraded or vanished due to disorder.
3. Physics-Agnostic Optimization: The loss function relies solely on the geometric "step-ness" of the conductance curve rather than targeting specific quantized values. The fact that the algorithm naturally converges to integer quantized plateaus suggests the solution is a manifestation of the underlying QPC physics rather than an overfitting artifact.
4. Simulation-to-Experiment Transfer: The work successfully translates a numerical optimization protocol, developed with full control over disorder parameters, to a real-world experimental setting, achieving marked improvements in device performance.

Results

• Simulations: In disorder-free simulations, the optimized gate configuration produced sharper conductance steps compared to a simple sweep of average voltages. When random disorder was introduced, the optimized configuration successfully restored the conductance staircase, whereas the unoptimized (disorder-free) configuration failed. The algorithm converged to specific voltage configurations that compensated for the specific disorder realization.
• Experiment: In the physical device, the algorithm autonomously tuned the gate voltages starting from a non-optimized state. The unoptimized trace showed no step-like features. After optimization (specifically at iteration 51), the conductance trace exhibited clear, well-defined quantized plateaus. Interestingly, the algorithm found multiple distinct voltage configurations (e.g., at iterations 51 and 85) that both yielded excellent step behavior, suggesting the existence of multiple viable solutions in the parameter space, potentially due to device asymmetry or the mitigation of specific disorder features. The optimization found an optimum within approximately 460 measurements.

Significance and Claims
The paper claims that automated tuning using CMA-ES is a viable method for enhancing device functionality and mitigating the effects of disorder in quantum devices. The authors emphasize that this approach yields optimized tunings that would be difficult to obtain manually, particularly in the presence of unpredictable disorder.

The significance lies in the demonstration that a "blind" optimization algorithm—one that does not know the physics of quantized conductance or the nature of the disorder—can effectively navigate a complex, noisy parameter space to restore fundamental quantum transport properties. The authors note that while the current experiment was limited by device asymmetry (optimizing only 7 of 9 gates), the results validate the concept of in situ disorder mitigation. They suggest that future improvements could involve multi-layer gate structures to allow full control over all pixels and the exploration of voltage configurations that vary as a function of the sweeping parameter. The work positions machine learning not just as a characterization tool, but as an active mechanism for overcoming material imperfections in semiconductor quantum technologies.